Events
CS Colloquium: AI Seminar - Responsible Design of Information Access Systems
Speaker: Fernando Diaz
Location:
60 Fifth Avenue, Room 7th Floor Open Space
Videoconference link:
https://cimsnyu.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=a3c15c1c-eb6d-4a96-ae68-ae6600e0f334
Date: Friday, April 8, 2022
Information access systems such as search engines and
recommender systems mediate the interaction between people and
overwhelming repositories of consumable data, including web content,
music catalogs, and social media. The prevalence of information access
problems has led to the adoption of ranking-based search and
recommendation algorithms across a variety of online services, either as
a core feature or supporting technology. While effective in research
settings, when deployed in production environments, these algorithms can
surface a variety of unanticipated social harms—including the unfair
allocation of exposure, misinformation, and stereotype reinforcement.
This talk will introduce a research program on the responsible design of
information access systems focused on understanding the relationship
between algorithms, individuals, and society. In order to explore this
approach, I will present recent work on the measurement and mitigation
of unfairness in ranking systems. I will begin by discussing how
inherent properties in ranking tasks and their solutions result in
unequal effectiveness for both end users and content creators. Motivated
by these issues, I will then define the expected exposure metric, a new
evaluation measure based on user behavior models that generalizes
classic utility metrics so as to incorporate unfairness. In order to
mitigate unfairness in existing ranking algorithms, I will describe and
evaluate a stochastic algorithm that directly optimizes expected
exposure. Although grounded in information access, these results have
implications for more general ranking settings found in natural language
processing and machine learning. I will close by proposing future work
focused on deepening and broadening the field of responsible information
access.